A DEEP LEARNING APPROACH FOR CROP TYPE MAPPING BASED ON COMBINED TIME SERIES OF SATELLITE AND WEATHER DATA
Keywords: Earth Monitoring, Satellite Image Time Series, Weather Time Series, Crop Dataset, Deep Learning, Crop Type Mapping, Sentinel-2
Abstract. Global Earth Monitor (GEM, Horizon 2020) takes advantage of the large volumes of available Earth Observation (EO), weather, climate and other non-EO data to establish economically viable continuous monitoring of the Earth. Within the GEM framework, the development of scalable and cost-effective solutions is being tested on several use-cases, with crop identification being one of them.
Crop identification uses a combination of EO and weather data to enable automatic identification of crops. The use case supports operational decisions when managing crops and the monitoring of actual vs. planned or reported agricultural land use (e.g., Common Agricultural Policy monitoring). Satellite data and weather data come at very different temporal and spatial resolutions: Sentinel-2 constellation nominally provides an observation of a field every 5 days at 10 m spatial resolution, while weather data has continuous hourly time series at multi-km spatial resolution. We have designed ad-hoc routines to spatially aggregate satellite data at field level and to systematically compose layers of different time discretization series, so that each EO is associated with a complete time series (of opportune length) of weather variables at daily resolution. For each field, we extract the time series of the median over field pixels of Sentinel-2 L1C bands, cloud mask and cloud probability. For doing this we take advantage of Sentinel Hub's Statistical API (Sinergise, 2020), that enables the retrieval of statistics of band values and derived indices over a specified geographic area and time range. Using meteoblue dataset API (meteoblue, 2017), complete time series of daily weather data (NEMS4 model, meteoblue, 2008) are then associated to each field observation, following the systematic layer composition approach mentioned above. An opportune time series length is defined for each of the 17 weather variables we considered. To handle this kind of multi-dimensional layered data, we use a flexible encoding-decoding framework (FlexMod, designed by TUM as part of GEM project): multiple encoders are designed for features of different time length (namely EO data and weather variables) and are then passed to the decoder via a mediator. Thanks to the flexible design of FlexMod framework, different models and architectures can be easily tested by simply defining new encoders and/or decoders. We present results obtained on a dataset in Slovenia, where crop fields are labelled according to a Hierarchical Crop and Agriculture Taxonomy (HCAT). This taxonomy, based on the EAGLE-Matrix and EU regulations, is the one adopted in the EuroCrops project (Schneider et al. 2021). The classification of field crops takes advantage of Sentinel-2 satellite data and Numerical Weather Prediction model output data. We exploit the potential of FlexMod to test different feature extractors, temporal encoding frameworks and decoders and we present a comparison between results obtained training a long-short term memory (LSTM) implementation (Breizhcrops, Rußwurm et al. 2020) and a Self-attention transformer model (Vaswani et al. 2017), the latter showing the best performances with accuracy 0.904 and Cohen’s kappa 0.824. We moreover investigate the role of weather data by benchmarking results against those obtained with just satellite imagery. To better appraise the influence of the weather data we analyse how perturbing weather data in the testing dataset affects the final results. So far, we obtain in both cases very similar accuracies and Cohen’s kappa. A deeper analysis of crop-specific scores (precision, recall, F1) suggests that the training and testing datasets are too limited in terms of size and crop variability to draw any general conclusion over the role of weather. As future developments, once the EuroCrops datasets are ready, we plan to expand the training and testing dataset to cover a higher variability of climatological areas and increase the numerosity of the so far under-represented crops, in the attempt to draw more general conclusions around the influence of weather and the predictability of specific crop classes. Moreover, given the encouraging scores, we aim to perform crop type mapping at least at European scale, thanks to the availability of the EuroCrops data and the cost-effective big data solutions developed during GEM project.